3 Data Points on Politeness: Does Saying “Please” Actually Hurt?

A look at token efficiency and direct commands.

I have a neighbor, let’s call him Bob. Bob is a retired contractor, the kind of guy who can fix a transmission with a paperclip but refuses to use online banking. Last summer, I helped him set up a basic smart speaker in his living room.

A week later, I stopped by for coffee. I watched Bob interact with the device.
“Alexa, please, if it’s not too much trouble, could you play that Credence Clearwater Revival song? The one I like?”

The machine paused, confused, and played a generic classic rock playlist. Bob looked defeated.

“It doesn’t listen,” he grumbled.

I looked at him. “Bob, it’s a database, not a waiter. Just say: Alexa. Play Fortunate Son.“

He tried it. It worked instantly.

This is the default stance for most people with a technical background. We view AI as a command-line interface (CLI). In a CLI, if you type sudo please apt-get update, the terminal laughs at you (or throws an error). We are trained to be terse, efficient, and brutal with our syntax.

But here is the reality check: LLMs are not command lines. They are probabilistic engines trained on human conversation. They predict the next word based on the “vibe” of the words that came before.

So, does being polite help? Or are you just burning tokens?